grpo_stable_reasoning_nolow_02kl_0714
GRPO RL checkpoints for a SimLingo-style driving VLA (InternVL2-1B + continuous speed-waypoint head).
Variant: nolow + KL 0.2 (0714 run). Companion run: grpo_stable_reasoning_nolow_nokl_0714 (KL off).
- Each
checkpoints/stepNNN.ckpt/pytorch_model.pt= consolidated fp32state_dict. .hydra/config.yaml= the training config — required to rebuild the model architecture for eval.
| file | step |
|---|---|
checkpoints/step100.ckpt/pytorch_model.pt |
100 |
checkpoints/step200.ckpt/pytorch_model.pt |
200 |
checkpoints/step300.ckpt/pytorch_model.pt |
300 |
checkpoints/step400.ckpt/pytorch_model.pt |
400 |
checkpoints/step500.ckpt/pytorch_model.pt |
500 |
import torch
sd = torch.load("checkpoints/step500.ckpt/pytorch_model.pt", map_location="cpu")
Note:
rollouts*.jsonl(RL training telemetry) is not included.